@dropdat/mcp

MCP server for dropdat. Lets any MCP-capable AI client (Claude Code, Cursor, Cline, Claude Desktop) recall, read, and save capsules in the user's dropdat library.

Tools

dropdat_recall calls the hybrid /capsules/search endpoint (vector + BM25 fused via RRF). Set OPENAI_API_KEY on the API for full semantic recall; without it the endpoint degrades to BM25.

Install

npm install -g @dropdat/mcp
# or, from source:
git clone https://github.com/dropdat/mcp.git
cd mcp && npm install && npm run build

Configure

Issue an API key in the dashboard → API Keys (token shown once, shape dk_live_…). Export it:

export DROPDAT_API_KEY=dk_live_xxx
export DROPDAT_API_BASE=http://localhost:8080   # or your prod URL

For local dev with DEV_AUTH_BYPASS=1 on the API, any non-empty DROPDAT_API_KEY value plus a JWT-shaped token will not work — the bypass uses DEV_USER_ID and ignores tokens. Use a real key against the running API.

Wire into a client

Claude Code

Add to ~/.claude/mcp.json (or project .mcp.json):

{
  "mcpServers": {
    "dropdat": {
      "command": "node",
      "args": ["/absolute/path/to/backend/mcp/dist/index.js"],
      "env": {
        "DROPDAT_API_KEY": "dk_live_xxx",
        "DROPDAT_API_BASE": "http://localhost:8080"
      }
    }
  }
}

Cursor / Cline / Claude Desktop

Same shape — point command/args at dist/index.js and pass the two env vars. Transport is stdio.

Develop

npm run dev   # tsx, no rebuild
npm run build # emit dist/

The compiled dist/index.js starts with a #!/usr/bin/env node shebang and is chmod +x'd on build — once published, clients can npx -y @dropdat/mcp instead of pinning a path.

Endpoint surface used

All against the Go API under /api/v1:

• GET /capsules?q=&tag=&limit=

• POST /capsules/search

• GET /capsules/{id}

• GET /capsules/{id}/lineage

• POST /capsules

Bearer auth — accepts either a Clerk session JWT or a dk_* API key. The MCP server uses the API-key path so it survives long-running agent sessions without a refresh dance.